skip to main content


Title: Evaluating BERT for natural language inference: A case study on the CommitmentBank
Natural language inference (NLI) datasets (e.g., MultiNLI) were collected by soliciting hypotheses for a given premise from annotators. Such data collection led to annotation artifacts: systems can identify the premise-hypothesis relationship without observing the premise (e.g., negation in hypothesis being indicative of contradiction). We address this problem by recasting the CommitmentBank for NLI, which contains items involving reasoning over the extent to which a speaker is committed to complements of clause-embedding verbs under entailment-canceling environments (conditional, negation, modal and question). Instead of being constructed to stand in certain relationships with the premise, hypotheses in the recast CommitmentBank are the complements of the clause-embedding verb in each premise, leading to no annotation artifacts in the hypothesis. A state-of-the-art BERT-based model performs well on the CommitmentBank with 85% F1. However analysis of model behavior shows that the BERT models still do not capture the full complexity of pragmatic reasoning, nor encode some of the linguistic generalizations, highlighting room for improvement.  more » « less
Award ID(s):
1845122
NSF-PAR ID:
10158557
Author(s) / Creator(s):
;
Date Published:
Journal Name:
Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)
Page Range / eLocation ID:
6085 to 6090
Format(s):
Medium: X
Sponsoring Org:
National Science Foundation
More Like this
  1. null (Ed.)
    Natural language inference (NLI) is an increasingly important task for natural language understanding, which requires one to infer whether a sentence entails another. However, the ability of NLI models to make pragmatic inferences remains understudied. We create an IMPlicature and PRESupposition diagnostic dataset (IMPPRES), consisting of 32K semi-automatically generated sentence pairs illustrating well-studied pragmatic inference types. We use IMPPRES to evaluate whether BERT, InferSent, and BOW NLI models trained on MultiNLI (Williams et al., 2018) learn to make pragmatic inferences. Although MultiNLI appears to contain very few pairs illustrating these inference types, we find that BERT learns to draw pragmatic inferences. It reliably treats scalar implicatures triggered by “some” as entailments. For some presupposition triggers like “only”, BERT reliably recognizes the presupposition as an entailment, even when the trigger is embedded under an entailment canceling operator like negation. BOW and InferSent show weaker evidence of pragmatic reasoning. We conclude that NLI training encourages models to learn some, but not all, pragmatic inferences. 
    more » « less
  2. One of the fundamental problems in Artificial Intelligence is to perform complex multi-hop logical reasoning over the facts captured by a knowledge graph (KG). This problem is challenging, because KGs can be massive and incomplete. Recent approaches embed KG entities in a low dimensional space and then use these embeddings to find the answer entities. However, it has been an outstanding challenge of how to handle arbitrary first-order logic (FOL) queries as present methods are limited to only a subset of FOL operators. In particular, the negation operator is not supported. An additional limitation of present methods is also that they cannot naturally model uncertainty. Here, we present BETAE, a probabilistic embedding framework for answering arbitrary FOL queries over KGs. BETAE is the first method that can handle a complete set of first-order logical operations: conjunction (∧), disjunction (∨), and negation (¬). A key insight of BETAE is to use probabilistic distributions with bounded support, specifically the Beta distribution, and embed queries/entities as distributions, which as a consequence allows us to also faithfully model uncertainty. Logical operations are performed in the embedding space by neural operators over the probabilistic embeddings. We demonstrate the performance of BETAE on answering arbitrary FOL queries on three large, incomplete KGs. While being more general, BETAE also increases relative performance by up to 25.4% over the current state-of-the-art KG reasoning methods that can only handle conjunctive queries without negation. 
    more » « less
  3. To build robust question answering systems, we need the ability to verify whether answers to questions are truly correct, not just “good enough” in the context of imperfect QA datasets. We explore the use of natural language inference (NLI) as a way to achieve this goal, as NLI inherently requires the premise (document context) to contain all necessary information to support the hypothesis (proposed answer to the question). We leverage large pre-trained models and recent prior datasets to construct powerful question conversion and decontextualization modules, which can reformulate QA instances as premise-hypothesis pairs with very high reliability. Then, by combining standard NLI datasets with NLI examples automatically derived from QA training data, we can train NLI models to evaluate QA models’ proposed answers. We show that our approach improves the confidence estimation of a QA model across different domains, evaluated in a selective QA setting. Careful manual analysis over the predictions of our NLI model shows that it can further identify cases where the QA model produces the right answer for the wrong reason, i.e., when the answer sentence cannot address all aspects of the question. 
    more » « less
  4. Abstract We investigate the link between individual differences in science reasoning skills and mock jurors’ deliberation behavior; specifically, how much they talk about the scientific evidence presented in a complicated, ecologically valid case during deliberation. Consistent with our preregistered hypothesis, mock jurors strong in scientific reasoning discussed the scientific evidence more during deliberation than those with weaker science reasoning skills. Summary With increasing frequency, legal disputes involve complex scientific information (Faigman et al., 2014; Federal Judicial Center, 2011; National Research Council, 2009). Yet people often have trouble consuming scientific information effectively (McAuliff et al., 2009; National Science Board, 2014; Resnick et al., 2016). Individual differences in reasoning styles and skills can affect how people comprehend complex evidence (e.g., Hans, Kaye, Dann, Farley, Alberston, 2011; McAuliff & Kovera, 2008). Recently, scholars have highlighted the importance of studying group deliberation contexts as well as individual decision contexts (Salerno & Diamond, 2010; Kovera, 2017). If individual differences influence how jurors understand scientific evidence, it invites questions about how these individual differences may affect the way jurors discuss science during group deliberations. The purpose of the current study was to examine how individual differences in the way people process scientific information affects the extent to which jurors discuss scientific evidence during deliberations. Methods We preregistered the data collection plan, sample size, and hypotheses on the Open Science Framework. Jury-eligible community participants (303 jurors across 50 juries) from Phoenix, AZ (Mage=37.4, SD=16.9; 58.8% female; 51.5% White, 23.7% Latinx, 9.9% African-American, 4.3% Asian) were paid $55 for a 3-hour mock jury study. Participants completed a set of individual questionnaires related to science reasoning skills and attitudes toward science prior to watching a 45-minute mock armed-robbery trial. The trial included various pieces of evidence and testimony, including forensic experts testifying about mitochondrial DNA evidence (mtDNA; based on Hans et al. 2011 materials). Participants were then given 45 minutes to deliberate. The deliberations were video recorded and transcribed to text for analysis. We analyzed the deliberation content for discussions related to the scientific evidence presented during trial. We hypothesized that those with stronger scientific and numeric reasoning skills, higher need for cognition, and more positive views towards science would discuss scientific evidence more than their counterparts during deliberation. Measures We measured Attitudes Toward Science (ATS) with indices of scientific promise and scientific reservations (Hans et al., 2011; originally developed by the National Science Board, 2004; 2006). We used Drummond and Fischhoff’s (2015) Scientific Reasoning Scale (SRS) to measure scientific reasoning skills. Weller et al.’s (2012) Numeracy Scale (WNS) measured proficiency in reasoning with quantitative information. The NFC-Short Form (Cacioppo et al., 1984) measured need for cognition. Coding We identified verbal utterances related to the scientific evidence presented in court. For instance, references to DNA evidence in general (e.g. nuclear DNA being more conclusive than mtDNA), the database that was used to compare the DNA sample (e.g. the database size, how representative it was), exclusion rates (e.g. how many other people could not be excluded as a possible match), and the forensic DNA experts (e.g. how credible they were perceived). We used word count to operationalize the extent to which each juror discussed scientific information. First we calculated the total word count for each complete jury deliberation transcript. Based on the above coding scheme we determined the number of words each juror spent discussing scientific information. To compare across juries, we wanted to account for the differing length of deliberation; thus, we calculated each juror’s scientific deliberation word count as a proportion of their jury’s total word count. Results On average, jurors discussed the science for about 4% of their total deliberation (SD=4%, range 0-22%). We regressed proportion of the deliberation jurors spend discussing scientific information on the four individual difference measures (i.e., SRS, NFC, WNS, ATS). Using the adjusted R-squared, the measures significantly accounted for 5.5% of the variability in scientific information deliberation discussion, SE=0.04, F(4, 199)=3.93, p=0.004. When controlling for all other variables in the model, the Scientific Reasoning Scale was the only measure that remained significant, b=0.003, SE=0.001, t(203)=2.02, p=0.045. To analyze how much variability each measure accounted for, we performed a stepwise regression, with NFC entered at step 1, ATS entered at step 2, WNS entered at step 3, and SRS entered at step 4. At step 1, NFC accounted for 2.4% of the variability, F(1, 202)=5.95, p=0.02. At step 2, ATS did not significantly account for any additional variability. At step 3, WNS accounted for an additional 2.4% of variability, ΔF(1, 200)=5.02, p=0.03. Finally, at step 4, SRS significantly accounted for an additional 1.9% of variability in scientific information discussion, ΔF(1, 199)=4.06, p=0.045, total adjusted R-squared of 0.055. Discussion This study provides additional support for previous findings that scientific reasoning skills affect the way jurors comprehend and use scientific evidence. It expands on previous findings by suggesting that these individual differences also impact the way scientific evidence is discussed during juror deliberations. In addition, this study advances the literature by identifying Scientific Reasoning Skills as a potentially more robust explanatory individual differences variable than more well-studied constructs like Need for Cognition in jury research. Our next steps for this research, which we plan to present at AP-LS as part of this presentation, incudes further analysis of the deliberation content (e.g., not just the mention of, but the accuracy of the references to scientific evidence in discussion). We are currently coding this data with a software program called Noldus Observer XT, which will allow us to present more sophisticated results from this data during the presentation. Learning Objective: Participants will be able to describe how individual differences in scientific reasoning skills affect how much jurors discuss scientific evidence during deliberation. 
    more » « less
  5. We investigate neg(ation)-raising inferences, wherein negation on a predicate can be interpreted as though in that predicate’s subordinate clause. To do this, we collect a largescale dataset of neg-raising judgments for effectively all English clause-embedding verbs and develop a model to jointly induce the semantic types of verbs and their subordinate clauses and the relationship of these types to neg-raising inferences. We find that some neg-raising inferences are attributable to properties of particular predicates, while others are attributable to subordinate clause structure. 
    more » « less